[Paper Review] Unsupervised Cross-domain Image Classification by Distance Metric Guided Feature Alignment

Qingjie Meng, Daniel Rueckert+1 more|arXiv (Cornell University)|Aug 19, 2020

Domain Adaptation and Few-Shot Learning35 references4 citations

TL;DR

This paper proposes Distance Metric Guided Feature Alignment (MetFA), a novel unsupervised domain adaptation method that learns discriminative and domain-invariant features without domain adversarial training. By jointly optimizing metric learning and class distribution alignment, MetFA achieves state-of-the-art performance in cross-device fetal ultrasound image classification, significantly improving target domain F1-score to 0.7874 compared to prior methods.

ABSTRACT

Learning deep neural networks that are generalizable across different domains remains a challenge due to the problem of domain shift. Unsupervised domain adaptation is a promising avenue which transfers knowledge from a source domain to a target domain without using any labels in the target domain. Contemporary techniques focus on extracting domain-invariant features using domain adversarial training. However, these techniques neglect to learn discriminative class boundaries in the latent representation space on a target domain and yield limited adaptation performance. To address this problem, we propose distance metric guided feature alignment (MetFA) to extract discriminative as well as domain-invariant features on both source and target domains. The proposed MetFA method explicitly and directly learns the latent representation without using domain adversarial training. Our model integrates class distribution alignment to transfer semantic knowledge from a source domain to a target domain. We evaluate the proposed method on fetal ultrasound datasets for cross-device image classification. Experimental results demonstrate that the proposed method outperforms the state-of-the-art and enables model generalization.

Motivation & Objective

• To address domain shift in medical imaging caused by differences in image acquisition devices, which limits model generalization.
• To overcome the limitations of domain adversarial training in unsupervised domain adaptation, which often fails to preserve discriminative class boundaries in the target domain.
• To learn a shared latent representation space that is both domain-invariant and discriminative across source and target domains.
• To transfer semantic knowledge from a labeled source domain to an unlabeled target domain through joint class distribution and metric learning.
• To enable effective, generalizable deep learning models for large-scale fetal ultrasound screening without requiring extensive target domain annotations.

Proposed method

• Proposes Distance Metric Guided Feature Alignment (MetFA), which explicitly learns a shared latent space without using domain adversarial training.
• Integrates metric learning to minimize intra-class variance and maximize inter-class variance across domains using a contrastive loss formulation.
• Employs class distribution alignment to align the predicted class distributions between source and target domains in the latent space.
• Uses a fixed prior distribution to embed features into a shared latent space before estimating domain-invariant prototypes.
• Combines metric learning with class distribution alignment in an end-to-end deep learning framework trained on source-labeled and target-unlabeled data.
• Applies t-SNE visualization and confusion matrices to analyze feature clustering and classification performance.

Experimental results

Research questions

• RQ1 Can metric learning be effectively used to learn domain-invariant and discriminative features without domain adversarial training?
• RQ2 Does joint optimization of metric learning and class distribution alignment improve cross-domain generalization in unsupervised domain adaptation?
• RQ3 How does the proposed MetFA method compare to state-of-the-art methods in cross-device fetal ultrasound image classification?
• RQ4 To what extent does the method preserve performance in the source domain while improving adaptation in the target domain?
• RQ5 Can the method generalize to real-world medical imaging applications with limited annotations and significant domain shift?

Key findings

• UDA-MetFA-V achieves an F1-score of 0.7874 on the target domain (device B) when device A is the source, outperforming the previous state-of-the-art MME method (0.4565).
• The method reduces confusion between similar anatomical views—such as LVOT and RVOT—demonstrating improved discriminative feature learning.
• t-SNE visualization confirms that features from the target domain are more aligned with source domain features and better separated by class.
• When device B is the source and device A the target, UDA-MetFA-V achieves an F1-score of 0.6303 on the target, significantly outperforming DANN (0.3450) and MME (0.2101).
• The model maintains strong performance on the source domain (F1-score of 0.7101), indicating that adaptation does not degrade source domain performance.
• Ablation studies show that the full UDA-MetFA-V model, combining metric learning and class distribution alignment, yields the best performance, confirming the effectiveness of the joint optimization strategy.